Better performance when sample timesteps is smaller？

[xiaotingxuan]

Hi,I am a greenhorn in diffusion model
I find something strange when I use diffusion prior model to generate image embedding.
First , I set prior_cond_scale = 2. and sample timesteps =64 , the cosine similarity between generated image embedding and target image embedding is around 0.68.
Then, I set prior_cond_scale = 1. the cosine similarity up to 0.75 .
I also find that the cosine similarity is higher if I decrease sample timesteps.
when I set prior_cond_scale = 1. sample timesteps=5, the cosine similarity can get around 0.80
Is this norm？ can anyone explain this for me?
(PS: I am sorry if my presentation is hard to understand，I am not a native English speaker.)

[funmooya]

how to change sample [timesteps]? I am not a native English speaker too.